Renewable Maoist Belt

Kishore Gawande (Texas A&M)

Devesh Kapur (University of Pennsylvania)

Shanker Satyanath (NYU)

ABSTRACT

Is there a causal relationship between shocks to renewable natural resources, such as agricultural and forest lands, and the intensity of conflict? In this paper we conduct a rigorous econometric analysis of a civil conflict that the Indian Prime Minister has called the single biggest internal security challenge ever faced by his country, the so called Maoist conflict. We focus on over-time within-district variation in the intensity of conflict in the states where this conflict is primarily located. Using a novel dataset of killings we find that adverse renewable resource shocks have a robust, significant association with the intensity of conflict. A one standard deviation decrease in our measure of renewable resources increases killings by 12.5% contemporaneously, 9.7% after a year, and 42.2% after two years. Our instrumental variables strategy allows us to interpret these findings in a causal manner.

 

1    

RENEWABLE NATURAL RESOURCE SHOCKS AND CONFLICT IN INDIA S RED BELT

(11,780 WORDS)

1. Introduction

The question of whether there is a causal relationship between shocks to renewable

natural resources (such as vegetation or water) and the intensity of conflict is an important one

for students of international political economy. 1 In recent years several IPE scholars have made

the claim that there is such a causal link.2 The claim is that where people are dependent on such

resources for their livelihoods, adverse shocks to these resources generate incentives to fight for

survival.3 However the basis for a causal link between renewable resource shocks and conflict

intensity has been strongly disputed, most powerfully by Gleditsch.4 As Schwartz et al.

environment-conflict research is methodologically unsound and fails to qualify as systematic

research. 5 Among his major criticisms are a failure to address reverse causality and a failure to

systematically control for alternative explanations for conflict.6 Some methodological

improvements have been made in a recent study by Theisen but fundamental concerns about

endogeneity raised by Gleditsch remain unaddressed in this literature.7

We attempt to address the serious concerns about reverse causality and omitted variables

bias that are currently obstacles to accepting the potential presence of a causal relationship by

                                                                                                                     1 This literature differs from the literature on the resource curse, in that the latter focuses on nonrenewable resources such as oil and minerals. 2 For example Homer-Dixon 1994; Hauge and Ellingsen 1998; Homer-Dixon 1999; Kahl 2006. 3 Mildner et al. 2011, 157. 4 Gleditsch 1998. 5 Schwartz et al. 2000, 78. 6 Schwartz et al. 2000, 78. 7 Theisen 2011.

2    

adopting an instrumental variables approach, and by focusing on over-time variation within

districts of a single large country. In doing so we are fully aware of the trade-off between tight

identification and generalizability that comes from shifting from a cross national to a single

country design, but in the absence of powerful cross-national instruments for renewable resource

shocks we consider this trade off to be essential to establish causality. A single country focus

also serves identification because the background national context cannot be the source of

omitted variables concerns.

We conduct a rigorous econometric analysis of a civil conflict that the Indian Prime

Minister has referred to as our

8 the so-called Naxal conflict. The Naxal conflict has largely been concentrated in a

corridor of states in the East-Central part of the country, and takes the form of a Maoist

insurrection with the goal of toppling the Indian government.9 According to official Indian

government data the conflict has resulted in 7862 deaths in the period 2000-2009.10 Whereas the

principal sources of politically motivated violence in India have either declined markedly (the

insurgencies in Kashmir and in Indi -East) or remained unchanged at relatively low

                                                                                                                     8The Economist 2/25/10. 9 The antecedents of the movement go back to a Maoist ideology inspired violent insurrection in Naxalbari, a village in the Eastern state of West Bengal in May 1967 (Chakravarti 2008). Violence spread over the next few years but it was violently put down by government forces and petered out by early

-eastern, central and eastern part of the country and gradually gained strength from the early 1990s when the many splintered groups b

India Marxist- he Maoist Communist Centre of India (MCCI), on September 21, 2004, the PGA was renamed as the PLGA. For a detailed account of the origins of the Naxalite movement see Banerjee (1980); Singh (1995) provides an empathetic account of the movement until the early 1990s from the vantage point of a senior police officer. 10 Data collated from the Annual Reports of Ministry of Home Affairs, Government of India.

3    

levels (communal violence) over the past decade, violence related to the left wing Naxal

(Maoist) movement remains the exception to this rule (Figure 1).

[Figure 1 here]

The Naxal conflict is well suited to the purpose of examining the causal links between

renewable resource shocks and conflict, because it meets one of the central conditions under

which the above link would plausibly hold; as we show later, large proportion of the population

in the Naxal belt of states is dependent on natural resources for its livelihood. A growing

literature has sought to understand the determinants of the Naxal conflict. However, aside from

methodological issues, this literature suffers from a serious problem. It relies on data from the

Incidents Tracking System (WITS) which measure conflict based on the study of reports in the

English language press. This is quite problematic given the limited coverage of English

language newspapers, the urban bias of the English language press, and the largely rural nature

of the conflict. Moreover the SATP data is only from 2005. We address this gap in the literature

on the Naxal conflict by creating our own conflict dataset which is additionally based on a

survey of the local language press, which has far superior rural coverage. As such we are able to

capture a substantially larger number of casualties than other currently available datasets.

As mentioned, our paper aims to explain within-district variation over-time in conflict

intensity. This focus allows us to use demanding specifications with both district and year fixed

effects, which effectively control for time invariant unobservable factors that may make districts

different from each other, as well as time variant shocks that may strike all districts at any given

point in time. Consistent with our goal of analyzing within-district over-time variation our core

analysis focuses on the four core states of the Naxal belt, where there is substantial year to year

4    

variation in killing (at the district level) to be explained. (The core states are Bihar, Jharkhand,

Chhattisgarh, and Andhra Pradesh.11) We instrument for a plausible proxy for over time

variations in renewable resources, and can thus make causal claims relating renewable resource

shocks to conflict intensity. Specifically, we use a satellite derived measure of vegetation

oxy for the agricultural and forest related natural resources on which a large

proportion of residents of the Naxal belt rely for their livelihoods. Since vegetation is plausibly

endogenous to conflict we instrument for vegetation with rainfall and find this to be a strong

instrument.12

Our main finding is that adverse vegetation shocks have a robust, significant association

with the intensity of conflict. A one standard deviation decrease in our measure of reweable

natural resources increases killings by 12.5% contemporaneously, 9.7% after a year and 42.2%

after two years. Our instrumental variables strategy allows us to interpret these findings in a

causal manner and thus contribute to the debate on the causal relationship between renewable

resource shocks and conflict.

In the next section of the paper we summarize the status of the literature on the

relationship between renewable resource shocks and conflict and then describe the findings of

the emerging literature on the Naxal conflict. In Section 3 we describe our causal story. The

data, which are a novel aspect of our paper, and the econometric strategy, are the subject of

Section 4. In Section 5 we present our main results and a range of robustness checks. Section 6

concludes.

2. Literature Review                                                                                                                      11 We also check for robustness when a fifth state Orissa is added. 12 Brown (2010) uses the same NDVI vegetation index that we use when analyzing conflict in Darfur, but does not instrument for vegetation.

5    

2.1. The debate over renewable resource shocks and conflict

Mildner et al. provide a comprehensive review of the debate over the causal relationship

between natural resources and conflict.13 Our aim here is to identify those contributions to the

debate that are specifically the concerns of this paper. Mildner et al. point out that the literature

can be divided into two broad categories. One focuses on the effects of renewable resource

deprivation on conflict, and another focuses on the effects of the resource curse. Our paper falls

in the category of the former set of concerns. Several scholars who have focused on the former

concern have emphasized the threats presented by depletion of croplands, forests, water, and fish

14 They have argued that such adverse resource shocks can cause

conflict by endangering livelihoods and forcing people to fight for their survival. Whereas

authors of this school primarily rely on case studies to justify their claims, their broad claims

have also found support in some statistical studies, for instance by Hauge and Ellingsen.15

The views espoused by this school have been criticized on several grounds. For instance,

Goldstone has pointed the conceptual shortcomings in the literature, notably the failure to define

variables clearly, as well as the failure to take account of alternative non-environmental

explanations for conflict.16 Especially relevant to our paper is Gleditsch criticism from a

methodological perspective.17 In his view, case-based scholars are being inadequately

systematic in their research. In particular, he has pointed out numerous failures in addressing

reverse causality and in rigorously controlling for alternative explanations for conflict.

                                                                                                                     13  Mildner et al. 2011.  14  Homer-Dixon 1994 and 1999 and Kahl 2006.  15  Hauge and Ellingsen 1998.  16  Goldstone 2001.  17  Gleditsch 1998.  

6    

Theisen has attempted to re

quantitative research design.18

natural resource shocks on conflict are likely exaggerated. Unfortunately Theisen too does not

have a compelling strategy for addressing reverse causation and his study is marred by the

inclusion of endogenous variables on the right hand side.

Since renewable , our paper is also related to

the cross country literature on the causal relationship between income and conflict.19 Most

relevant is the paper by Miguel, Satyanath, and Sergenti which uses rainfall as an instrument for

per capita GDP when studying civil conflict in Africa.20 The big difference is that we are

focusing on a level of analysis (district level) that is bereft of GDP statistics. In using a

vegetation measure our paper offers a novel alternative way of addressing economic welfare at a

local level in environments where citizens are reliant on agriculture/forest products for their

livelihood.

2.2. The literature on the Naxal conflict

While India has had a history of radical peasant movements most observers trace the

an anti-landlord peasant uprising in Naxalbari, a village in West Bengal. 21 While today in media

and popular discussion, the term

Maoism can be seen as the one of most radical brands of Naxalism. The 2004 merger of the

(Maoist),

along with the CPI (ML)                                                                                                                      18  Theisen 2011.  19  Collier and Hoeffler 1998, Fearon and Laitin 2003.  20  Miguel, Satyanath, and Sergenti 2004.  21  Singha Roy 2004.

7    

doctrine, the specific ideology of Maoism can be found in the official documents of the CPI

-

(LWE).22

the Scheduled

Castes (or Dalits), Scheduled Tribes (Adivasis) and Muslims the tribal population has gained

least and lost most in post-independence India, both because they are spatially isolated and their

fragmentation has made them unable to articulate their grievances through the democratic and

electoral process.23 Their major problems are land alienation, denial of forest rights, and

displacement by development projects and national parks and sanctuaries. The British introduced

laws which independent India unquestioningly inherited, turning the stewards of the forest into

subjects of the forest department. The resulting failures of the state and of the formal political

system, according to Guha, have provided a space for the Maoists.

The magnitude of the problem has spawned a growing academic literature. Barooah

examined which socio-economic variables explain the existence of Naxalite activity in some

districts of India but not others using data from the Indian Planning Commission and South

Asian Intelligence Review.24 The dependent variable is the likelihood of violence. The main

findings are that the probability of a district being Naxalite-affected rises with an increase in its

poverty rate and falls with a rise in its literacy rate and that Naxalite activity in a district reduces

the overall level of violent crime and crimes against women. These results are from cross

sectional OLS regressions at the district level, and pool across heterogeneous regions.

                                                                                                                     22 There are several careful ethnographic studies of the Maoist movement. For Bihar see Bhatia (2005) and for Jharkhand see Shah (2010). Harriss (2010) provides a good review of some of these studies. 23  Guha 2007.  24  Barooah 2008.  

8    

Iyer examines terrorist incidents in general, where terrorist activity includes separatist

movements, communal violence, and the Naxal insurgency.25 Using data from Global Terrorism

Database 2, Rand-MIPT Terrors Database, and from Planning Commission and South Asian

Intelligence Review, Iyer finds correlation between violence and poverty. Like Barooah, her

findings are based on cross sectional OLS regressions at the district level.

Sen and Teitelbaum trace the history of the Maoist movement and use the World

Incidents Tracking System (WITS) data on Maoist violence to examine the effects of mining.26

They conclude that the geographical spread of Maoist movement is simply too wide to be

accounted for mainly by mining activity. However, the weaknesses of their data on violence as

well as mining activity, and failure to address endogeneity are potentially serious shortcomings.

Endogenity is also an issue for Hoelscher, Miklian and Vadlamannati who analyze cross-

sectional data from six Indian states - Chhattisgarh, Andhra Pradesh, Orissa, Jharkhand, Bihar

and West Bengal from 2004 to 2010.27 Their data combines that from the South Asian Terrorism

(WITS) and the Global Terrorism Database. Using probit and negative binomial estimation

techniques they find that conflict increases with forest cover, prevalence of conflict in

neighboring districts, and the population share of members of scheduled castes and tribes.

Gomes combines databases from different sources on Maoist incidents/violence into one

database.28 His paper looks at landholdings and historical land institutions, and finds a strong

effect of land inequality on Maoist violence. As described above this paper purports to be a

                                                                                                                     25  Iyer 2009.  26  Sen and Teitelbaum 2010.  27  Hoelscher, Miklian, and Vadlamannati 2011.  28  Gomes 2011.  

9    

district level analysis but does not include district fixed effects and also does not address

endogeneity.

Van den Eynde examines the strategic choices of targets and the intensity of violence of

Maoist insurgents using data from the South Asia Terrorism Portal (SATP) between 2005 and

2010.29 Using variation in annual rainfall in a panel of district level casualty numbers he finds

that negative labor income shocks increase violence against civilians to prevent them from being

recruited as police informers. It also increases the number of rebel attacks against the

government, but only if the rebels' tax base is sufficiently independent from local labor

productivity (in this case access to key mineral resources). The casualty data used in the paper is

from the SATP which, as mentioned, relies primarily on the English language press.

In a more general way our paper contributes to the statistical literature on the

determinants of violence in India, for instance the work by Wilkinson on ethnic riots in India and

by Jha and Wilkinson on prior military training on violence during the partition of British

India.30 Our paper adds to this literature by examining the effects of changes in renewable

natural resources (namely vegetation) on violence.

3. Causal story relating renewable resource shocks to intensity of conflict

Our causal claim is that over-time, within-district variations in the intensity of conflict in

the Maoist belt of states are significantly driven by shocks to renewable natural resources.31 For

this claim to be plausible we must first establish that the livelihoods of people in the region are

significantly reliant on renewable natural resources. We present several facts that establish this

                                                                                                                     29  Van den Eynde 2011.  30  Wilkinson 2004 and Jha and Wilkinson 2012.  31 As mentioned the core of this belt in our analysis is the four states listed at the outset of the paper where 90% of Maoist related killings occur.

10    

basic point.

The two major renewable resources on which people in the region rely for their

livelihood are agricultural land and forest land. More than two-thirds of the employment in the

Maoist belt is in agriculture. In 2004 (the mid-point of the time period of our analysis), 73.6% of

agricultural land will plausibly affect economic welfare of numerous residents of the region.

Aside from agricultural land, a significant proportion of the population in the region

relies on forests for its livelihood. The Maoist region includes a significant number of Scheduled

Tribes.32 In two of the four states in our analysis the share of tribals in the population exceeds

25%.33 These tribals tend to live in or in the proximity of forests. In Chhattisgarh almost half of

the land area in tribal districts consists of forest cover; in Jharkhand it is almost a third.34 Tribals

have close cultural and economic links with the forest and depend on forests for a significant part

of their subsistence and cash livelihoods, which they earn from fuel wood, fodder, poles, and a

range of non-timber forest products, such as fruits, flowers, medicinal plants and especially

tendu leaves (used to wrap tobacco flakes and make beeris, a local Indian hand-rolled cigarette).

While dozens of press and academic accounts of the tribal peoples in central and eastern

                                                                                                                     32 The nomenclature goes back to the 19th Century when the British expansion into the forested and hilly terrain of central India encountered culturally distinct groups living in relative geographical isolation. The first Census in 1872 categorized these coScheduled Districts Act. The Indian Constitution in 1950 listed them in a "Schedule", granting them special

33 Indian Planning Commission, http://planningcommission.nic.in/data/datatable/0904/tab_119.pdf 34 Forest Service of India. http://www.fsi.nic.in/sfr2003/forestcover.pdf. 60% of the forest cover of the country and 63% of the dense forests lie in 187 tribal districts (GOI, 2008).

11    

India stress their dependence on forest resources for their livelihoods, a single macro-quantitative

estimate of the degree of dependence is not available. We provide here the evidence that is

available, which is strongly suggestive of heavy dependence. A World Bank study has found

that in the state of Jharkhand fuel wood from forests supplies an average of 86 percent of energy

needs while fodder from the forest provides about 55 percent of input requirements for domestic

livestock for communities living proximate to forests.35 The FAO cites a study of tribal

households in Orissa which found that an average tribal family drew about one-half of its annual

income from forests, 18 percent from agriculture, 13 percent from cattle and 18 percent from

other employment.36 A study of forest fringe communities in the Jhabua district of Madhya

Pradesh (which is close to the Maoist belt) measured specific components of household income

and subsequent dependence on natural resources (including forests), and found this dependence

to be three-fourths for those in the lowest 25% income quartile and nearly two-thirds for those in

the 25-50 percent income quartile.37 Another study (in West Bengal which also neighbors the

Maoist belt) found that the share of forest income in total income for tribal households ranges

from 55-86%.38

None of the above is intended to suggest that the bulk of tribals rely on the forest to the

exclusion of agriculture for their livelihoods. In his landmark work on the tribes of India, the

ethnologist von Fürer-Haimendorf has categorized tribes into three categories based on their

primary economic activity: food gatherers and hunters, shifting cultivators and settled farm

                                                                                                                     35  World Bank 2006.  36http://www.fao.org/docrep/x2450e/x2450e0c.htm#joint%20forest%20management%20in%20india%20 and%20the%20impact%20of%20state%20control%20over%20non%20wood%20f 37   .  38  Das 2010.  

12    

populations.39 What is clear from this categorization is that tribals are dependent on either the

forest and/or agricultural resources for their livelihoods. This includes cases where tribals

dependence on forest and agricultural resources is seasonal; a widespread pattern noted by von

Fürer-Haimendorf is one where the tribals rely on forest products during much of the first half of

the year, engage in subsistence agriculture during much of the second half, and supplement their

income with activities like basket weaving during the monsoon season.

To sum up, we believe it is reasonable to expect adverse shocks to the renewable

resources of forests and agricultural land to substantially affect the livelihoods of most residents

of the Maoist belt of states. The question then is, what are our expectations about how this would

result in greater intensity of conflict?

A shock to livelihoods creates incentives for people to join the Maoists as an alternative

means of survival. However, we do not expect this to be immediately reflected in more intense

conflict. Accounts by journalists (and there are no academic studies) stress the Maoist strategy of

gradualism when inserting a new recruit into an actual combat situation.40 Recruitment is

followed by training. Recruits are taken far away from their home villages and are sometimes

even sent for training to other parts of India, such as the North East, before returning to fight on

familiar turf. This means that the full effects of resource shock driven recruitment on intensity of

conflict should not be observed contemporaneously, but with a lag.

Another reason to expect a lagged effect is that the ability of Maoists to successfully fight

the government is enhanced by support from the local population. One potential source of

support is goodwill resulting from some genuine service being provided by the Maoists. The

                                                                                                                     39  Fürer-Haimendorf 1982.  40  Chakravarti 2008.  

13    

immediate reaction of many to an adverse economic shock is to borrow from a moneylender.

Moneylenders in the Maoist belt are well known for being ruthless in demanding collateral when

repayment is not forthcoming.41 One of the major services that Maoist cadres provide is to offer

protection for villagers from confiscation of assets by moneylenders.42 This protection likely to

be needed and provided, not when the loan is first taken, but rather when repayment is due. The

goodwill of the local population should thus be at its highest not contemporaneously with the

adverse economic shock, but after the passage of some time. This would generate incentives for

Maoists to launch their most furious attacks at a lag from the adverse economic shock.

4. Econometric Strategy and Data

4.1 Econometric Strategy

The first aspect of our econometric strategy is the choice of a variable to capture

renewable resource shocks in the Naxal belt. In the context of the Naxal belt such a variable

should pick up differences between good and bad agricultural years and the differences between

years when the forest is relatively fecund and when it is not. Both of these goals are plausibly

achieved with a measure that captures the density of vegetation. Less dense vegetation should

both be associated with poorer crops and with a less rich forest environment. A time varying

measure of vegetation should thus pick up shocks to the main natural resources which residents

of the Naxal Belt rely on for their livelihoods. Therefore, our proxy for natural resources in the

Maoist region is the degree of vegetation in a district-year.

Vegetation may, of course, be endogenous to conflict. For instance the Indian

government may destroy forests in places where conflict is anticipated, in order to facilitate

                                                                                                                     41  Guha 1999  42 Pandita 2011.

14    

counter insurgency operations. Thus it is necessary to instrument for vegetation. We use

rainfall as an instrument for vegetation, and this critical step allows us to make causal claims

about the effect of natural resource shocks on conflict intensity.

In our tables we present results with and without instrumentation. The uninstrumented

regressions have the following structure, with i referring to the district and t referring to the year:

ln(TotalDeaths)i,t = i,t-1 1Vegetationi,t 2Vegetation i,t-1 3Vegetation i,t-2

+ Year fixed-effects + District fixed-effects + ei,t, (1)

where, after accounting for the year and district fixed effects, the error term ei,t is identically and

independently distributed normally. All our models are estimated with robust standard errors

that are clustered at the district level. Since district dummies are included in the model, the

coefficient estimates are based on the within-district variation in the data. The year-fixed effects

account for macro-events that impact all districts in a given year, for example inflation. This is a

demanding specification that takes account of observable and unobservable time invariant

differences between districts (for example the share of Scheduled Tribes in a district varies very

little over time and is thus absorbed by the district dummies). It also takes account of all common

shocks affecting all districts in the sample at any point in time.

In our instrumental variables regressions we instrument for the vegetation variables

above with three lags of rainfall in addition to contemporary rainfall. Given that there was a

change towards a much more highly defined sensor to capture vegetation in 2001 our analysis

begins with this year and runs to 2008 (the last year for which we have complete killings data.)

15    

We also conduct robustness checks using the Arellano and Bond GMM technique in which

lagged levels serve as instruments for lagged first differences. More details are provided in the

results section.

4.2 Data

Killings

The primary data innovation in the paper is our attempt at improving on four other

sources of

from the Ministry of Home Affairs (MHA); the RAND-MITP Terrorism Incident database; data

from the Worldwide Incidents Tracking System (WITS) from National CounterTerrorism Centre

and data from the South Asian Terrorism Portal (SATP). The Ministry of Home Affairs compiles

data on Naxal violence, dividing it into two categories the number of casualties (deaths) and

the number of violent incidents. While this data is available from 2001 onwards it is available

only at the state level. The Rand-MIPT and the WITS data sets are world-wide and their Indian

data is ad hoc and do a poor job at capturing Indian data reported by the non-English language

press. The SATP data set has been assembled by the Institute for Conflict Management, available

at their website, the South Asia Terrorism Portal. These data are based primarily on reports in the

major English press, but does a considerably better job than the other two, especially in more

recent years. However, this data set is only from 2005 and even after that it is not comprehensive

while ours goes back to 2000.

The data set we have assembled does not draw from secondary sources. Instead for each

state we examined reports from at least four distinct media sources:

i) National English press

16    

ii) The local (or state) edition of a national English daily (if present)

iii) Local language (vernacular) dailies (at least two, except in Andhra Pradesh where we

used one, Eeenadu, the largest circulation Telugu language daily in that state.

iv) Two wire services:Press Trust of India (PTI) and India Abroad News Service (IANS).

The data sources used to construct this data base of all Maoist incidents is detailed in Appendix

Table 1. Further details about the dataset, including coding procedures, are provided in our

online appendix (not for publication, but submitted separately for reviewers at the end of the

tables).

[Appendix Table A1 here]

Each incident is geocoded at the district level. Our core measure of total casualties in

incidents involving Maoists sums the number of civilian deaths, security personnel deaths and

Maoist deaths for each year for each district. (We also look at these separately.) We are able to

capture a substantially larger number of casualties than the widely used South Asian Terrorist

portal. The extent to which other databases underestimate the actual number of deaths is evident

in Figure 2. Only the SATP data is collected at the district level, but as Figure 2 indicates, it

underreports considerably compared to our dataset.

[Figure 2 here]

There are two years in which our data picks up fewer killings than the aggregate data

reported by the Government of India MHA numbers. Given the capacity and incentive problems

of MHA data (in some cases states that reported more violence got more money to fight the

Maoists), as a benchmark this data itself should be taken with caution. Is it possible that killings

are over-reported by the vernacular press? This would imply that somehow the language medium

17    

would result in over-reporting. This might be if one believes that the rural press has empathy

with the Maoists. But the newspapers we have used in this study both English and vernacular

are all owned by urban capitalists, some regional capitalists and some national (none of the news

sources are communist affiliated which might lead to an exaggeration of deaths). Moreover, such

an argument implies that those who read/write English are less biased than those who do not,

which has little basis.

We focus our attention in this paper to all districts in the four states of Andhra Pradesh,

Bihar, Chh

with the Maoist movement. 90% of total killings in our national data over the 2001-2008 period

are from these districts.

The number of states and districts has changed over time. Chhattisgarh was carved out of

Madhya Pradesh in 2000, and Jharkhand out of Bihar in 2000. To keep the unit of observation

consistent across the years, we map districts into a common set of regions using the concordance

constructed by Kumar and Somanathan.43 Specifically, we use their 1991 mapping of over 600

Indian districts into 404 regions. For the four red states that are the focus of this study, we have

data on 68 regions over the 2000-2008 period. (For purposes of avoiding confusion we refer to

these regions as districts everywhere else in this paper.)

Rainfall

Rainfall data are from the high resolution (1° ×1° latitude/longitude) gridded daily

rainfall dataset for the Indian region, kept annually since 1951 by the India Meteorological

Department (IMD). The daily rainfall data are archived at the National Data Centre, IMD Pune.

Rajeevan et al. of IMD Pune describe the collection of IMD data as follows: IMD operates about                                                                                                                      43  Kumar and Somanathan (2009, Tables 4-8).  

18    

537 observatories, which measure and report rainfall that has occurred in the past 24 h ending

0830 h Indian Standard Time (0300 UTC).44 In addition, most of the state governments also

maintain rain gauges for real-time rainfall monitoring. IMD has the rainfall records of 6329

stations with varying periods. Out of these, 537 are IMD observatory stations, 522 are under the

Hydrometeorology program and 70 are Agromet stations. The remaining are rainfall-reporting

stations maintained by state governments. Rajeevan et al. show that this data correlates well

both spatially and inter-temporally with the VASClimo dataset, is a global gridded rainfall

dataset constructed in Germany.

We match the capital cities of (districts in) the 68 regions in the four red states to nearest

rain station and ascribe that rain data to the district.45 The daily rainfall data are aggregated into

annual data.

Vegetation

Satellite imagery is now widely used in the sciences to track changes in vegetation and

forest cover.46 We use the normalized difference vegetation index (NDVI) to measure annual

change in vegetation for Indian districts. The NDVI data are derived from visible infrared and

near-infrared data acquired from the MODIS sensor (Moderate Resolution Imaging

Spectroradiometer) on NASA satellites. The NDVI index is computed as NDVI = (NIR VIS) /

(NIR + VIS), where NIR is the Near Infrared Band value and VIS is the visible light or the Red

                                                                                                                     44  Rajeevan et al. 2006.  45 Where many districts map into one region, the median district mapped to the nearest rain station. 46 For example Myneni et al. 1998; Tucker et al. 2001; Nemani et al. 2003. In the Indian context, for example, Panigrahy et al. (2010) find dense forests denuding at 0.72% per annum in the Western Ghats, while Prabhakar et al. (2006) measure deforestation in the Himalayas.

19    

Band value recorded by the satellite sensor.47 The NASA site explains the computation as

follows: healthy vegetation absorbs most of the visible light that hits it and reflects a large

portion of the near-infrared light. Unhealthy or sparse vegetation reflects more visible light and

less near-infrared light. NDVI for a given grid ranges from 1 to +1. Zero means no vegetation

while values above 0.5 indicate substantial density of green leaves. NDVI is a good measure of

density from tree rings.48

We mapped the data from MODIS into 1° × 1° latitude/longitude grids for India. For

each grid, monthly NDVI data are averaged to obtain the mean annual NDVI over the period

2001-2009.49 They are then mapped into Indian regions, specifically into the 68 red districts as

was done for the rainfall data.

Other Data

For robustness, we also report results using consumption data. The source for our

consumption expenditures data is the Annual National Sample Surveys (NSS) taken over the

2000-2009 period are. Monthly per capita expenditure (MPCE) at the household level collected

in these surveys are averaged using sample-proportionate-to-population weights to obtain MPCE

                                                                                                                     47 http://earthobservatory.nasa.gov/Features/MeasuringVegetation/measuring_vegetation_4.php 48   . al. 2000.  49 MODIS data are a substantial improvement over its predecessor, the AVHRR sensor which provided 20 years of data going back to 1980. According to Hansen et al (2003), MODIS is a significant advance due to three reasons. First is the finer instantaneous field of view of MODIS (250 and 500 m2) as compared to AVHRR instruments (1 km2). Second, MODIS was built with seven bands specifically designed for land cover monitoring, allowing for greater mapping accuracy due to more robust spectral signatures. It also reduces background scattering from adjacent pixels as the MODIS land bands were designed to limit the impact of atmospheric scattering. Third, 500-m red and near-infrared data, two bands important for land cover mapping, are created from averaged 250-m imagery. This resampling reduces the percent contribution of adjacency effects on 500-m pixels for these bands, allowing for improved land cover estimates. The result is a dataset that reveals far more spatial detail than previous efforts. For our purpose the great differences between the coarser-scale maps in the pre-MODIS era prevents a comparison on pre-2001 vegetation data with the MODIS data. It is a primary reason for focusing our analysis to the 2001-2008 period even though we have killings data in the pre-2000 years.

20    

measure for each of the 68 districts annually over the period. The accuracy of our MPCE

calculations was verified by matching our figures with those in the NSS summary reports (at the

state level) produced by the government of India. While the consumption measure affirms our

main results, we choose not to emphasize the consumption results due some technical

measurement issues that have been raised by experts. The Tendulkar Commission report

expands upon the difficulties of measuring poverty consistently over time from the NSS data.50

Translating an objective poverty baseline measure based on calorific intake into a measure based

on consumption spending in rupees is fraught with problems. Changes in the price of food and

measurement error in changes in the quantity and quality of food consumed lead to measurement

error in measuring the poverty line, based on which other measures of poverty are calculated,

such as poverty headcount. Since our econometric analysis is primarily based on within-district

variation in the data, measurement error in every period is less acceptable. We also want to drive

the conflict literature to emphasize local knowledge and local circumstance much more than it

has done in the past, as is reflected in our use of vegetation.

The literature has attempted to attribute the Maoist conflict to a number of sources

including the extent of mining activity in this mineral-rich region of India, the predominance of

underprivileged persons, specifically tribals (Scheduled Tribe or ST) and of people of lower

caste (scheduled caste or SC) in these regions, the possibility of spillovers effects from

neighboring regions, and the extent of inequality. We have made every attempt to incorporate

these influences in our analysis. We measure mining of bauxite and iron ore in each district

annually from Indiastat.com and compiled using the Indian Mineral Year Book annual district

production figures published by the Indian Bureau o

                                                                                                                     50  Tendulkar et al. 2009.  

21    

population that is SC and ST (from the NSS), a consumption expenditure Gini at the district

level, and the number of the two closest districts that have experienced Maoist violence in the

past year to capture spillover effect.

In sum, we have assembled a comprehensive data set for the 68 districts in the four red

states of Andhra Pradesh, Bihar, Chattisgarh and Jharkhand annually over 2001-08. Since we

incorporate dynamics and use up to three lags, our results are based on 340 observations

consisting of a balanced panel of the 68 districts over the period 2004-08. Table A2 in the

appendix provides panel descriptive statistics for the variables we use in our analysis.

[Appendix Table A2 here]

5. Results

In Table 1 we examine the relationship between vegetation and total deaths (our core

measure of the intensity of conflict). Column 1 shows that vegetation is negatively associated

with total deaths and the relationship is significant at the 1% level. The z statistic at the bottom

1 2 3 =0. If

vegetation and its lags are strongly correlated, their collective significance is relevant. The sum

of the coefficients is an essential input into calculating the long-run impact of a vegetation shock.

Rejecting the null hypothesis, as the z statistic implies, indicates a strong negative (long-run)

association of vegetation with killings. This is the first indication that vegetation is negatively

associated with the intensity of conflict.

[Table 1 here]

Since our dependent variable is a count variable it is worth checking if the OLS result

22    

changes when we apply a negative binomial model.51 Since the NB model encompasses the

Poisson model (which assumes that the conditional mean of killings equals its conditional

tests for over-dispersion and rejects the Poisson process for killings. Column 2 shows the results

with the NB model are substantively unchanged.

Column 3 offers another robustness check, this time with an Error Correction Model.

Although the ECM has traditionally been used for time-series analysis of cointegrated data

series, De Boef an

cointegrated (but stationary, as the Maoist deaths and the NDVI series are in our case). 52

Equation (1) is easily translated into the ECM form as:

i,t = ( -1)ln(TotalDeaths)i,t-1 i,t-1

i,t i,t-1 + Year fixed-effects + District fixed-effects + e i,t, (2)

where i,t = X i,t X i,t-1. The ECM regresses the first difference

of Maoist-related deaths on one lag of Maoist-related deaths, one lag of vegetation, the first

difference of vegetation, the first difference of lagged vegetation, and district and year fixed

effects. The immediate impact of a 1- Maoist-

related deaths. The long-run impact of a 1-

change in Maoist-related deaths. The third column in Table 1 reports estimates from the ECM.

                                                                                                                     51 The link function in the NB model is log-linear, so the OLS model in logs produces coefficients comparable to those from the NB model. 52  De Boef and Keele 2008.  

23    

The immediate short-run impact of a negative vegetation shock of 0.01 is estimated to increase

Total Deaths by 13.99 × 0.01 = 14% in the distributed lag model (1) and by 14.90 × 0.01 =

14.9% in the ECM. The long run impact of the same shock is estimated to be (0.01) × ( 13.99

7.86 7.83) / (1 0.08) or a 32.3% increase in Total Deaths using (1), and (0.01) × ( 29.67 /

(.920) equal to (the same) 32.3% increase in Total Deaths using the ECM. 53

Time series scholars are concerned that panels with short T (less than 15) may be

afflicted with Nickell bias when fixed effects and the lagged dependent variable are

simultaneously included in the same specification.54 In columns 4 and 5 we thus check if the

results change when the lagged dependent variable is dropped from the column 1 and 2

specifications and find that there is no substantive difference. The Arellano Bond GMM

approach offers an alternative technique to address Nickell bias. Our results are robust to the use

of the Arellano-Bond technique (not shown).55

Vegetation may be endogenous to conflict, for example if the police clear forests for ease

of fighting. If shocks to killings are negatively correlated with vegetation because forests are

denuded to facilitate counter-insurgency, then the negative coefficients on vegetation in Table 1

may be the result of this endogeneity bias. In order to address this concern we resort to an

instrumental variables strategy, using rainfall as an instrument for vegetation. Table 2 shows the

strong relationship between rainfall and vegetation, affirming the case for rainfall as a good

instrument for vegetation. The Kleibergen-Paap Wald statistic of 10.19 indicates no weak

instruments problem. Specifically, the K-P statistic implies the (small-sample) bias in the 2-

                                                                                                                     53 The Akaike Information Criterion justifies the use of a model with two lags (results available on request). 54  Nickell 1981.  55 The Arellano-Bond results are available from the authors. They mirror the IV results we present. Even though the Arellano-Bond models are in first differences, and hence capture changes in the variables, their similar results are a strong robustness check of our core OLS-IV and NB-IV results.

24    

stage estimates is less than 5% of the bias in the corresponding OLS estimates in Table 1.56 As

the table shows, there is generally a positive association between lagged rainfall and vegetation.

The only significant negative coefficients are for forward variables --contemporaneous and once

lagged rain in the twice lagged rain regression -- which is explained by mean reversion because

of which we see good rain years following bad rain years.

[Table 2 here]

Table 3 shows our core instrumental variables second stage results.57 In column 1 we

report the results for a linear instrumental variables regression (OLS-IV). The strong negative

association between vegetation and total deaths remains even with instrumentation. The point

estimates are much larger, which is consistent with our instruments addressing measurement

error with respect to vegetation. Essentially, the IV results exploit only exogenous variation in

vegetation, which allow us to make causal statements. In column 2 we undertake the same

exercise, but with a negative binomial model to better account for the count nature of our

dependent variable. The results remain substantively unchanged, and are stronger in the same

direction as the OLS-IV results.

[Table 3 here]

The strong effect of twice lagged vegetation on killing is consistent with the dynamic

causal story described in Section 3. An adverse vegetation shock gives Naxal leaders the

opportunity to recruit fighters into training camps. The recruitment is not overnight. Maoist

cadres come in and undertake a variety of actions from providing rudimentary medical services

to income augmenting schemes like offering protection against money lenders and forest

                                                                                                                     56  Stock and Yogo 2004.  57All estimation is done using Stata 11. The 2SLS results use code due to Baum et al. 2007.

25    

officials.58 This means that the full effects of recruitment and training are not observed

contemporaneously with the adverse vegetation shock, but rather after a lag. The long-run result,

however, is a large increase in Maoist-related deaths.

The quantitative implications of the IV estimates bear this out. The OLS-IV results

indicate that an adverse shock that reduces vegetation by 0.011 (a one within-standard deviation

change see table A2) increases killings by 12.5% contemporaneously, 9.7% after a year and

42.2% after two years. While the contemporaneous and the first lag effects are measured

imprecisely, the second lag effect is statistically significant at 1%. The sum of the three

coefficients is statistically significant at 1% (indicated by the z statistic towards the bottom). The

long-run impact of an adverse shock that reduces NDVI by a one (within) standard deviation in a

single period, is an increase of 64% in total deaths. While not reported, an error correction model

(ECM) confirms the results.59

The estimates from the NB model in column 2 are notable as well. A one within-

standard deviation decrease in NDVI increases killings by 16.5% contemporaneously, 27.7%

after a one-year lag and 76.1% after a two-year lag. Columns 3 and 4 of Table 3 show that the

results are substantively unchanged if we drop the lagged dependent variable from the column 1

and 2 specifications (to address Nickell bias concerns).

Instrument diagnostics are reported at the bottom of table 3. The Hansen J-statistic

cannot reject the joint null hypothesis that (i) the instruments are uncorrelated with the error

                                                                                                                     58  Pandita 2011.  59 We estimate the ECM it in two stages, first predicting the three vegetation variables, and then estimating the ECM in (2) after replacing the lagged and differenced vegetation variables with their instrumented counterparts. The estimates from the ECM replicate the immediate and long-term impact of a vegetation shock in Table 3. The second lag of vegetation is statistically significant, confirming its importance. The long-run impact of a negative vegetation

58.37 / (.923) = 63.2%. The ECM results are available from the authors.

26    

term, and (ii) the instruments (rainfall) are correctly excluded from the estimated equation. The

Anderson-Rubin (A-R) statistic tests the significance of the endogenous regressors. The statistic

rejects the joint null hypothesis that the coefficients of the endogenous regressors in the

structural equation (the three vegetation measures) are jointly equal to zero.60 This statistic is

important in assessing the results when instruments are weak, as we will see below. The

rejection of the null is not surprising here since we expect rainfall to be a good instrument for

vegetation.

What is the effect of successive years of vegetation depletion on violence in Maoist

areas? This is a relevant question to ask because press reports indicate that construction, mining,

and population pressures have all contributed to the degradation of forest and agricultural land in

India. We investigate the impact of vegetation depletion by interacting the three lagged

vegetation variables Vegetationt, Vegetationt-1, and Vegetationt-2. The interactions are highly

correlated with the individual lagged measures, and including a full set of variables produces

collinear models whose coefficient estimates are not individually informative.61 In the last three

columns of Table 3 we present results from the set of interactions that are informative. The

vegetation interactions are instrumented with the three rainfall lags and their interactions.

                                                                                                                     60 The AR statistic actually tests the joint null hypothesis that (i) the coefficients of the endogenous regressors in the structural equation are each equal to zero, and that (ii) the overidentifying restrictions are valid. Thus, the AR

1 2 3=0) or because the instrument orthogonality (with the error) conditions fail. A variety of tests have been proposed (Kleibergen 2002; Moreira 2003) that maintain the hypothesis that the instruments are valid. Moreira shows that the AR test is optimal when the equation is just identified (number of excluded instruments equals the number of endogenous variables), which is the case here. A caveat is that with extremely weak instruments and strong endogeneity of the regressors, the power of the A-R test (its ability to not falsely reject) may be poor. 61 The partial correlation (after partialing out district and year fixed effects) of Vegetationt and the interaction (Vegetationt-1, ×Vegetationt-1), for example, is over 0.75. Including all possible interactions and the linear terms terms leads to large coefficients and standard errors, with opposite signs.

27    

If the interactions have negative coefficients, it implies that successive years of depletion only

worsen violence. The first column shows that Vegetationt-2 and its interaction with Vegetationt

and Vegetationt-1have negative coefficients, although none of them are statistically significant.

This is likely due to the collinearity among the three variables.62 Dropping the linear term

Vegetationt-2, in the next columns yields negative signs on both interaction terms and a

statistically significant estimate on Vegetationt-2 × Vegetationt-1. This implies that the marginal

effect of an increase in 2-period lagged vegetation, or Total D Vegetationt-2, equals

( 36.89 × Vegetationt) (84.83 × Vegetationt-1). Evaluated at their means, the marginal effect is

49.94, implying that a one-sd increase (of 0.011) in 2-period lagged vegetation decreases total

killings by 55%.

The last column shows results from one-lag interactions of Vegetationt-2 with

Vegetationt-1 and Vegetationt-1 with Vegetationt. The negative and statistically significant

coefficient on Vegetationt-2 × Vegetationt-1 indicates that the impact on killings due to the

denuding of vegetation two periods ago (t-2) is strongly exacerbated if the vegetation is not

renewed in the following year (t-1). The marginal effect t-2, equals

45.92 and is precisely measured. It affirms the finding above.

The causal story we favor is the standard one in the resource shock-conflict literature, in

which adverse resource shocks reduce the opportunity cost of fighting for rebels. Given this

causal path it is important to check that the results are not driven by killings by security forces

rather than rebels. To check this we disaggregate total deaths into civilian casualties, Maoist

(militants) casualties, and security (police) casualties. Table 4 presents instrumental variables

results for these categories using the same IV strategy as for the first two columns of Table 3.

                                                                                                                     62 They are, however, informative about the marginal effect (see below).

28    

The negative relationship between vegetation and casualties is in evidence for all three types of

casualties. We also do not see a robust difference in the responses to vegetation of killings of

rebels vs. security troops. (The slight difference in timing in the OLS-IV does not carry over

into the negative binomial IV regressions.) Our first-hand experience with gathering killings

data indicates civilian victims are sometimes classified as Maoists (especially when the incident

is reported by the security forces). Combining the two categories yields very similar results.63

[Table 4 here]

One potential concern is whether the vegetation results reflect aspects of the technology

of fighting? Less vegetation could change the balance of power between security forces and

rebels by making it harder for rebels to conduct insurgency operations from the forest because it

is harder for them to remain hidden as they attack or retreat, and therefore in conducting guerilla

operations. In addition, thinner vegetation should make it easier to conduct for security forces to

conduct counterinsurgency operations and kill rebels because the latter will find it harder to hide.

If either of these mechanisms is driving the results, vegetation should be positively associated

with security deaths minus rebel deaths. The OLS-

4 show that this is not the case and this increases our confidence that our

vegetation results are likely not picking aspects of the technology of fighting.64

We have mentioned the measurement issues that affect household consumption

expenditure data collected annually by the national sample surveys (NSS). Regardless, we think

it is useful to check whether our core results are robust to using monthly per capita consumption

expenditures (MPCE) as measures of livelihood and poverty in place of vegetation. The first

                                                                                                                     63 These results are available from the authors. 64 There is no reason to run a negative binomial model here because the left hand side can take values of less than zero.

29    

four columns of Table 5 present coefficient estimates corresponding to the non-IV estimates in

Table 1. The results are substantively very similar and support inferences from the vegetation

results. The ECM indicates that the long-run impact of an adverse shock of 136 rupees

(approximately one within-sd) to monthly income increase Maoist-related deaths by ( 1.515 ×

0.136 = 23.3%. We take this as an affirmation of our main findings from the

vegetation data that poverty is inversely correlated with Maoist-related killings.

[Table 5 here]

The IV results are reported in the last two columns of Table 5. The K-P statistics indicate

that the consumption measure instruments weakly with rainfall (which is consistent with the

measurement error critique in the Tendulkar Commission report mentioned earlier). However,

the Anderson-Rubin p-value gives us a way of producing estimates that take account of the

weak-instrument problem.65 While the individual coefficients on consumption are insignificant

the Anderson-Rubin p-value of .008 confirms that the variables are jointly robust even after

standard errors are adjusted to take account of the extent of weakness of the instruments.

We now consider the influence of several key control variables on our core regressions.

These variables have appeared in other studies and in popular stories and reports of the conflict.

Mining contracts sold by the state to firms for bauxite and iron ore mining that rampantly denude

forest lands and displaces its denizens, have been blamed as a source of conflict. Tribals and the

lower caste population have been especially affected by such actions, and districts with greater                                                                                                                      65 The small Kleibergen-Paap statistics in Table 3 indicate that rainfall is a weak instrument for consumption. The problem of weak instruments has deservedly attracted much attention, for strong instruments are hard to find. Weak- -R), seek to produce valid inference

about endogenous regressors and confidence intervals for their coefficients despite the weakness of instruments (see also Kleibergen 2002 and Moriera 2003). Technically, the Anderson-Rubin statistic AR defines is a quadratic form that is a function of the IV estimates b: q(b) = AR. Inverting this function produces interval for b. If the intervals preserve the signs for b then they are weak-instrument robust estimates. The A-R statistic of 3.55 rejects the null hypothesis that the coefficients on the three instrumented consumption variables are jointly zero, at the 1% level.

30    

proportion of SC and ST populations are likely to experience greater displacement and therefore

propensity to violence as they are recruited into the Maoist fold. Income inequality may be a

uniting force among insurgents and districts with greater inequality have been shown to

experience greater intensity of conflic .66 Finally, the incidence of

conflict in neighboring districts may spill over into conflict in the present district.

[Table 6 here]

Table 6 examines the robustness of our primary results to the inclusion of these

influences as control variables. The first five columns include each of these controls, one

variable at a time in the OLS-IV models. The spillover effect is statistically significant and

indicates that if one of the closest two neighboring districts experiences Maoist violence, then

killings increase 26.4%. While there is evidence that SC, inequality and mining all increase

killings, that evidence is weak since the coefficients are statistically not significant. The sixth

column pools these controls into one specification. Two results are notable. The first is that core

results are unchanged with the addition of these controls. Second, the proportion of the ST

population is positively associated with more deaths. An increase of .01 in the tribal population

(on account of migration, for example), results in a 1.606% increase in killings.

[Table 7 here]

Could our results be driven by states in which Scheduled Tribes have not achieved

representation in the set of parties that are competing in state politics? The states where there is

such representation are Jharkhand and Chhattisgarh, which were carved out as separate states in

November 2000 precisely to ensure greater representation of tribals. Ironically the Maoist

insurgency surged in these areas after the new states were created and the tribals got more                                                                                                                      66  Nepal, Bohara and Gawande 2011.  

31    

political representation. In Table 7 we interact the vegetation variables and state dummies, with

Andhra Pradesh as the baseline (excluded) state. The z statistics at the bottom summarize the

results by summing the coefficients for each state across the three coefficients for each state

tal

z-statistics for the OLS_IV regression indicates that the total effect of vegetation is negative in

all states. The only state for which the total effect of vegetation is insignificant in the OLS-IV

regression is Chhattisgarh. (The z-statistic in the Negative Binomial regression is positive, albeit

insignificant.) We verified that the effect for Chhattisgarh is significantly lower than for

Jharkhand where representation of Scheduled Tribe is highest. This suggests that lack of

political representation is likely not the reason for the differences across states.

There is a plausible explanation for why conflict intensity in Chattisgarh is less sensitive

to vegetation shocks than in other states in the Naxal Belt. This is the one state where potential

Maoist recruits have been deliberately moved by the government, away from forest and

agricultural lands and into guarded camps along the highway. Since the effect of this measure is

that the livelihoods of numerous potential recruits in Chhattisgarh are no longer subject to

variations in vegetation, it makes sense that one does not observe a strong vegetation-killings

link in this state.

While not affecting our claims with respect to the hard core Naxal Belt states it

interesting to assess the effects of slightly expanding the definition of the Naxal Belt. The most

logical expansion is to Orissa because, while not traditionally considered to be part of the Naxal

Belt, parts of Orissa were subject to Naxal conflict. We coded killings in Orissa and checked if

32    

our core Table 3 results were affected by the inclusion of Orissa. We found that our core results

were unchanged (not shown).

Finally, could the results be driven by violations of the exclusion restriction? One

possible violation highlighted by Sarsons in the context of riots all over India lies in the need to

account for the building of large dams in our analysis.67 This may be a relevant consideration for

consideration for our analysis because there is no over-time variation for our region of analysis

in the building of large dams. (Using the standard definition of 50 meters, no new large dams

were built in our region after 2000.)

Another exclusion-restriction-related consideration is that fighters may not want to fight

when it is raining, so rainfall may be directly affecting conflict via this mechanism. This

violation is hard to measure. Whether or not this biases against our results and whether the bias is

large enough to kill our result thus remains an unknown. However, we do note that this is not

likely to be a serious factor affecting our analysis because our dataset indicates that the most

killings occur outside of the rainfall season in good and bad rainfall years alike.

6. Conclusion

We demonstrate that even after addressing endogeneity concerns and subjecting our

findings to multiple robustness checks there is a strong, and substantively large relationship

between adverse resource and the intensity of conflict in Naxal belt. Aside from

contributing to a central debate in the civil conflict literature, our results carry implications for

policy. One approach would be to help citizens deal with adverse income shocks that result from                                                                                                                      67  Sarsons 2011.  

33    

adverse resource shocks. The creation of the National Rural Employment Guarantee Act

(NREGA) could offer promise as an insurance against income shocks. (NREGA only came into

effect midway through our period of analysis, so we cannot offer any firm statistical results along

this line.) Another possibility is implementing a catastrophe insurance scheme, when the income

shocks are spatially correlated, for instance due to extreme weather conditions. However, to the

extent that these options rely on the same government machinery that has, through acts of

commission, exacerbated the plight of tribals, their prospects may be questionable.

Our paper also suggests an alternative path that

technical capability. Giving tribals greater access to forests and a range of forest products,

whose consumption is the only available option during times of distress, can provide them with a

critical self-insurance mechanism. Tribals have been denied this access for many years,

especially after passage of the Forest Conservation Act of 1980. Reversing this situation was the

goal of the Scheduled Tribes and Other Traditional Forest Dwellers (Recognition of Forest

Rights) Act 2006, which came into force on 1st January 2008. Our paper suggests that rigorous

implementation of the government withdrawal called upon by this act may offer great promise.

34    

BIBLIOGRAPHY

between maximum latewood density of annual tree rings. and NDVI based estimates of forest

productivity. International Jopurnal of Remote Sensing 21: 2329 2336.

Brown, Ian A. 2010. Assessing Eco-Scarcity as a Cause of the Outbreak of Conflict in Darfur: A

Remote Sensing Approach. International Journal of Remote Sensing 31 (10), 2513-20.

ivreg2: Stata module for extended

instrumental variables/2SLS, GMM and AC/HAC, LIML, and k- Boston

College Department of Economics, Statistical Software Components S425401.Downloadable

from http://ideas.repec.org/c/boc/bocode/s425401.html.

Banerjee, Sumanta. 1980. In the wake of Naxalbari: A history of the Naxalite movement in India

(Calcutta: Subarnarekha, 1980).

Bhatia, Bela. 2005. The Naxalite Movement in Central Bihar. Economic and Political Weekly,

Vol.XL: 1536-43.

Bhattacharya, Prodyut, Lolita Pradhan and Ganesh Yadav. 2010. Joint forest management in

India: Experiences of two decades. Resources, Conservation and Recycling. 54: 469-480.

35    

Nepal, Mani, Alok K. Bohara and Kishore Gawande. 2011. More Inequality, More Killings: The

Maoist Insurgency in Nepal. American Journal of Political Science 55: 886 906.

Bohlken, Anjali Thomas and Ernest Sergenti. 2010. Economic Growth and Ethnic Violence: An

Empirical Investigation of Hindu Muslim Riots in India. Journal of Peace Research 47 (5), 589-

600.

Borooah, Vani K. 2008. Deprivation, Violence, and Conflict: An Analysis of Naxalite Activity in

the Districts of India International Journal of Conflict and Violence 2(2): 317-333.

Chakravarti, Sudeep. 2008. Red Sun: Travels in Naxalite Country. New Delhi: Penguin Viking

Press.

Oxford Economic

Papers 50.

Collier, Paul and Anke Hoeffler World Bank Policy

Research Paper 2355 (May).

Collier, Paul and Journal

of Conflict Resolution 46(1).

36    

Das, Nimai. 2010. Incidence of forest income on reduction of inequality: Evidence from forest

dependent households in milieu of joint forest management," Ecological Economics, Volume 69

(8): 1617 1625.

American Journal of

Political Science 52(1): 184-200.

Ev

Eynde, Oliver Vanden. 2011. Targets of Violence: Evidence from India's Naxalite Conflict.

http://personal.lse.ac.uk/vandeney/Targets_of_Violence.pdf

American

Political Science Review, 97(1): 75-90.

Gadgil and Guha, 1994. Ecological conflicts and the environmental movement in India

Development and Change, 25: 101-136.

Gallo, Kevin P.; Ji, Lei; Reed, Brad; Dwyer, John; and Eidenshink, Jeffrey. 2004. "Comparison

of MODIS and AVHRR 16-day normalized difference vegetation index composite data"

Geophysical Research Letters 31: L07502.

37    

Gleditsch, Nils-Petter. 2008. Armed Conflict and the Environment: A Critique of the Literature.

Journal of Peace Research 35 (3), 360-380.

Goldstone, Jack A.. 2001. Democracy, Environment, and Security in Environmental Conflict,

ed. Paul Diehl and Nils-Petter Gleditsch, 84-108. Boulder: Westview.

Gomes, Joseph Flavian. 2011. The Political Economy of the Maoist Conflict in India: An

Empirical Analysis. Available at:

http://www.uclouvain.be/cps/ucl/doc/core/documents/Gomes.pdf

http://planningcommission.gov.in/reports/publications/rep_dce.pdf

Guha, Ramachandra. 2007. Adivasis, Naxalites and Indian Democracy. Economic and Political

Weekly, August 11: 3305-3312.

Hansen, M. C., R. S. DeFries, J. R. G. Townshend, M. Carroll, C. Dimiceli, and R. A. Sohlberg.

2003. Global Percent Tree Cover at a Spatial Resolution of 500 Meters: First Results of the

MODIS Vegetation Continuous Fields Algorithm Earth Interactions 7, Paper No. 10: 1-15.

Hauge, Wenche and Tanja Ellingsen. 1998. Beyond Environmental Security: Causal Pathways

to Conflict. Journal of Peace Research 35 (3), 299-317.

Homer-Dixon, Thomas. 1994. Environmental Scarcities and Violent Conflict: Evidence from

Cases. International Security 19 (1), 5-40.

38    

Homer-Dixon, Thomas. 1999. Environment, Security, and Violence. Princeton: Princeton

University Press.

Harriss, John. 2010. The Naxalite/Maoist Movement in India: A Review of Recent Literature

ISAS Working Paper, No. 109 08 July.

Hidalgo, F. Daniel, Suresh Naidu, Simeon Nichter, and Neal Richardson.2010. Economic

Determinants of Land Invasions Review of Economics and Statistics, 92(3), 505-523.

Hoelscher, Kristian, Jason Miklian and Krishna Chaitanya Vadlamannati. 2011. Hearts and

Mines: A District-Level Analysis of the Maoist Conflict in India. Available at: http://www.uni-

heidelberg.de/md/awi/professuren/intwipol/india.pdf

Jenkins, Stephen P. and Philippe Van Kerm. 2009. The Measurement of Economic Inequality in

Brian Nolan, Wiermer Salverda and Tim Smeeding (eds.) Oxford Handbook on Economic

Inequality. New York: Oxford University Press: 40-71.

Jha, Saumitra and Steven Wilkinson. 2012. Veterans, Organizational Skill and Ethnic Cleansing:

Evidence from the Partition of South Asia

Stanford Graduate School of Business Working Paper No 2092.

Kahl, Colin. 2006. States, Security, and Civil Strife in the Developing World. Princeton:

Princeton University Press.

39    

Kleibergen, F. 2002. Pivotal statistics for testing structural parameters in instrumental variables

regression.Econometrica70: 1781 1803.

Kumar, Hemanshu andRohini Somanathan. 2009. Mapping Indian Districts Across Census

Years 1971-2001 Centre for Development Economics, Delhi School of Economics, Working

Paper 176.

Leamer, Edward E. 1978. Specification Searches: Ad Hoc Inference with Nonexperimental Data,

Wiley.

Mehta, J. and Venkatraman, S. 2000. Poverty Statistics: Bermicide's Feast, Economic and

Political Weekly, 35: 2377-2381.

Miguel, Edward, Shanker Satyanath, and Ernest Sergenti. 2004. Economic Shocks and Civil

Conflict: An Instrumental Variables Approach Journal of Political Economy, 112(4), 725-753.

Miguel, Edward and Shanker Satyanath. 2011. Re-examining Economics Shocks and Civil

Conflict. American Economic Journal, Applied Economics 3(4), October 2011.

Mildner, Stromy-Annila, Gitta Lauster, and Wiebke Wodny. 2011. Scarcity and Abundance

Revisited: A Literature Review on Natural Resources and Conflict. International Journal of

Conflict and Violence 5 (1), 155-172.

Ministry of Home Affairs. 2006. Annual Report 2005-2006 (New Delhi: Govt. of India).

Moreira, M. 2003. A conditional likelihood ratio test for structural models.Econometrica71:

1027 1048.

40    

Myneni, R., C. Tucker, G. Asrar, and C. Keeling.1998.Interannual variations in satellite-sensed

vegetation index data from 1981 to 1991, Journal of Geophysical Research 103(D6), 6145

6160.

Narain, Urvashi, Shreekant Gupta, and Klaas van Veld. 2005. Poverty and the Environment:

Exploring the Relationship between Household Incomes, Private Assets, and Natural Assets

Resources for the Future Discussion Paper 05-18.

Nemani, R. R., C. D. Keeling, H. Hashimoto, W. M. Jolly, S. C. Piper, C. J. Tucker, R. B.

Myneni, and S. W. Running. 2003. Climate-driven increases in global terrestrial net primary

production from 1982 to 1999, Science 300: 1560 1563.

Nickell, Stephen John. 1981. Biases in Dynamic Models with Fixed Effects. Econometrica 49

(6): 1417-26.

Nillesen, Eleanora and Philip Verwimp.2010. Grievance, Commodity Prices and Rainfall: A

Village Level Analysis of Rebel Recruitment in Burundi HiCN Working Paper 58.

http://www.hicn.org/papers/wp58.pdf

Panigrahy, Rabindra K.,; Kale, Manish P., Dutta, Upasana, Mishra, Asima, Banerjee, Bishwarup

Singh, Sarnam. 2010. Forest cover change detection of Western Ghats of Maharashtra using

satellite remote sensing based visual interpretation Current Science 98 (05)

41    

M. Poffenberger, B. McGean (Eds.). 1996. Village Voices, Forest Choices: Joint Forest

Management in India, Oxford University Press, New Delhi.

Prabhakar, R, Somanathan, E., Mehta, Bhupendra Singh. 2006. How degraded are Himalayan

forests? Current Science 91 (01).

M. Rajeevan, Jyoti Bhate, J. D. Kale and B. Lal. 2006. "High resolution daily gridded rainfall

data for the Indian region: Analysis of break and active monsoon spells." Current Science 91,

2006: 296-306.

Sarsons, Heather. 2011. Rainfall and Conflict.

Manuscript.http://www.econ.yale.edu/conference/neudc11/papers/paper_199.pdf

Schwartz, Daniel M., Tom Delgiannis, and Thomas Homer-Dixon. 2000. The Environment and

Environmental Change and Security Project Report, 6 (Summer), 77-94.

Sen, Rumela and Emmanuel Teitelbaum. 2010. Mass Mobilization a

Maoists http://web.gc.cuny.edu/dept/rbins/conferences/RBFpdf/Sen-TeitelbaumMaoists.pdf

Shah, Alpa. 2010. In the Shadows of the State: Indigenous Politics, Environmentalism and

Insurgency in Jharkhand, India. Durham and London: Duke University Press.

42    

Shah, Amita and Sajitha O.G. 2009.Dwindling forest resources and economic vulnerability

among tribal communities in a dry/sub-humid region in India. Journal of International

Development 21: 419 432

Singh, Prakash. 1995. The Naxalite Movement in India New Delhi: Rupa and Co.

SinghaRoy, Debal K.. 2004. Peasant Movements in Post-Colonial India: Dynamics of

Mobilisation and Identity New Delhi: Sage.

Stock, James H. and Motohiro

Identification and Inference in

Econometric Models: Essays in Honor of Thomas J. Rothenberg. Cambridge: Cambridge

University Press.

Paper.

Tendulkar, Suresh D., R. Radhakrishna, and Suranjan Sengupta. 2009. "Report of the Expert

Group to review the Methodology for Estimation of Poverty". Government of India Planning

Commission.

Theisen, Ole Magnus. 2008. Blood and Soil: Resource Scarcity and Armed Conflict Revisited.

Journal of Peace Research 45, 801-818.

43    

Tucker, C. J., D. A. Slayback, J. E. Pinzon, S. O. Los, R. B. Myneni, and M. G. Taylor. 2001.

Higher northern latitude normalized difference vegetation index and growing season trends from

1982 to 1999, International Journal of Biometeorology 45: 184 190.

Wilkinson, Steven 2004. Votes and Violence: Electoral Competition and Ethnic Riots in India.

Cambridge: Cambridge University Press.

-Dependent People in India.

Washington D.C.: World Bank. Report No. 34481-IN.

0

200

400

600

800

1000

1200

1400

2000 2001 2002 2003 2004 2005 2006 2007 2008

Num

bero

fDeaths

Year

Figure 2: Major Maoist States (AP BI JH CH)Comparison of CASI, MHA, SATP, WITS, and RAND Datasets

Deaths Related to Maoist Violence 2000 2008CASI

Ministry of Home Affairs

South Asian Terrorism Portal

Worldwide IncidentsTracking SystemRAND

ln(Total Deaths) Total Deaths ln(Total Deaths) ln(Total Deaths) Total DeathsOLS NB ECM OLS NB

ln(Total Deaths)t 1 0.0802 0.0944 0.920***[0.0693] [0.106] [0.069]

Vegetationt 13.99*** 21.87*** 13.87*** 21.75***[4.732] [7.215] [4.789] [7.257]

Vegetationt 1 7.857* 16.48** 29.67*** 8.234* 17.20**[4.565] [7.399] [10.01] [4.569] [7.287]

Vegetationt 2 7.827 14.33* 8.398 15.10*[5.567] [8.287] [5.701] [8.362]

Vegetationt 13.99***[4.731]

Vegetationt -1 7.827[5.567]

N 340 340 340 340 340R 2 0.13 0.205 0.502 0.125 0.205k 77 77 77 76 76

1.098*** 1.105***z ( VEGt ) 2.964*** 4.133*** - 2.98*** 4.232***

Note: 1. Robust standard errors in brackets. Clustered (by disctrict) in OLS models. *** p<0.01, ** p<0.05, * p<0.1 2. All models include district fixed effects and year dummies. NB models estimated with true fixed-effects (in the model). 3. NB models: Dependent variable is Total Deaths. Pseudo R 2 reported. Error correction model (ECM): Dependent variable is ln(Total Deaths). See Equation (2). 4. is the overdispersion parameter in NB model. > 0 indicates overdispersion; = 0 indicates Poisson is appropriate. 5. All coefficients are to be interpreted as in a log-linear model. 6. z ( t) tests the hypothesis: Vegetationt + Vegetationt 1 + Vegetationt 2 = 0. 7. Models without lagged dependent variables are reported to check for Nickell bias (Nickell 1981).

Table 1: Vegetation and Total Deaths from Maoist incidentsOLS and Negative Binomial (NB) models (Uninstrumented)

Dependent Variable:

Vegetationt Vegetationt -1 Vegetationt -2 Vegetationt Vegetationt -1 Vegetationt -2

ln(Total Deaths)t !1 0.0002 !0.001 !0.001

[0.001] [0.001] [0.001]

Raint !0.104 !0.0144 !0.652*** !0.108 !0.0017 !0.639***

[0.243] [0.247] [0.247] [0.243] [0.246] [0.246]

Raint !1 1.393*** 0.0828 !0.369* 1.396*** 0.0709 !0.381*

[0.211] [0.238] [0.216] [0.211] [0.239] [0.217]

Raint !2 0.272 1.585*** 0.0688 0.271 1.590*** 0.0733

[0.225] [0.223] [0.254] [0.224] [0.225] [0.253]

Raint !3 1.189*** 0.419 1.349*** 1.181*** 0.446* 1.376***

[0.254] [0.256] [0.253] [0.254] [0.253] [0.253]

District fixed-effects Yes Yes Yes Yes Yes Yes

Year dummies Yes Yes Yes Yes Yes Yes

N 340 340 340 340 340 340

R2

0.218 0.222 0.206 0.217 0.219 0.202

k 78 78 78 77 77 77

Weak Instrument Diagnosis:

Partial R2

0.14 0.134 0.158 0.14 0.135 0.162

First-stage F 12.86 13.61 14.55 12,89 13.42 15

Kleibergen-Paap (WI)

Note:

1. Robust standard errors clustered (by district). *** p<0.01, ** p<0.05, * p<0.1

2. Models include 68 district fixed effects and 5 year dummies.

Table 2: First stage for IV resuDependent variable: Lags of Vegetation Index (NDVI)

10.19 10.25

OLS-IV NB-IV OLS-IV NB-IV NB-IV NB-IV NB-IV

SECOND STAGE:

ln(Total Deaths)t !1 0.0398 0.104 0.070 0.0729 0.0793

[0.0780] [0.102] [0.101] [0.101] [0.101]

Vegetationt !11.54 !15.05 !11.35 !14.75

[14.28] [20.08] [14.20] [19.99]

Vegetationt !1 !8.831 !25.17 !8.905 !24.54

[13.31] [22.47] [13.31] [22.24]

Vegetationt !2 !38.45*** !69.15*** !39.16*** !70.97*** !27.79

[14.49] [23.60] [14.38] [23.76] [39.76]

Vegetationt !2 " Vegt !22.29 !36.89

[35.33] [33.27]

Vegetationt !2 " Vegt !1 !66.71 !84.83** !101.8***

[43.23] [33.70] [37.00]

Vegetationt " Vegt !1 !10.76

[32.33]

N 340 340 340 340 340 340 340

k 77 77 76 76 80 79 80

1.107*** 1.115*** 1.094*** 1.099*** 1.103***

z ( VEGt ) !3.276*** !3.647*** !3.335*** !3.629***

p -val ( VEGt ) 0.001 0.0003 0.001 0.0003

p -val (VEGt !VEGt -1) 0.896 0.761 0.906 0.763

p -val (VEGt !VEGt -2) 0.251 0.081 0.230 0.071

p -val (VEGt -1!VEGt -2) 0.173 0.245 0.167 0.219

ln(Total Deaths) / Vegt !2 !64.3*** !49.94*** !45.92***

[25.46] [11.80] [11.12]

ln(Total Deaths) / Vegt !4.413

[13.25]

FIRST STAGE:

#Instruments 4 4

Kleibergen-Paap (WI) 10.19 10.25

Hansen's J 0.060 0.084

Hansen's J (p- value) 0.807 0.772

Anderson-Rubin (A-R) 3.547 3.752Anderson-Rubin (p val) 0.008 0.005

Note:

1. Robust standard errors in brackets. Clustered (by district) in OLS models.*** p<0.01, ** p<0.05, * p<0.1

2. See Notes to Table 1

3. NB-IV models: Predicted endogenous variables in first stage used as regressors in the second stage.

4. ln(Deaths) / Vegetationt !2 reported at the means of other interactions.

5. p -val (VEGt !VEGt -1) is the p -value for the hypothesis test Vegetationt = Vegetationt -1

Table 3: Vegetation and Total Deaths from Maoist incidents

IV models. Dependent Variable : Total Deaths (logged for OLS-IV)

Vegetation Vegetation Interactions

ln(Civilian) ln(Maoist) ln(Security) ln(Sec)!ln(Mao) Civilian Maoist Security

Vegetationt !14.83 !1.648 !17.84** !16.63 !31.51 8.53 !54.78

[10.65] [12.06] [8.655] [14.05] [26.19] [25.52] [38.51]

Vegetationt !1 !14.25 !1.927 2.287 4.174 !53.31* 15.87 !25.39

[11.02] [11.01] [8.497] [11.36] [29.61] [27.58] [34.42]

Vegetationt !2 !22.79** !22.65* !12.77 9.682 !54.46** !72.94** !75.83**

[9.868] [12.45] [8.758] [14.39] [26.88] [30.62] [36.74]

ln(Civilian Deaths)t !1 !0.008 0.036

[0.094] [0.150]

ln(Maoist Deaths)t !1 !0.016 !0.082

[0.075] [0.122]

ln(Security Deaths)t !1 0.044 !0.128

[0.103] [0.198]

ln(Total Deaths)t !1 !0.0177

[0.074]

z ( VEGt ) !3.652*** !1.712* !2.494** !0.147 !3.518*** !1.272 !3.301***

p -val ( VEGt ) 0.0003 0.088 0.013 0.883 0.0004 0.203 0.001

p -val (VEGt !VEGt -2) 0.642 0.288 0.713 0.230 0.565 0.048 0.744

N 340 340 340 340 340 340 340

k 77 77 77 77 77 77 77

1.252*** 1.332*** 1.76***

#Instruments 4 4 4 4

Kleibergen-Paap 10.4 10.12 10.22 10.19

Hansen's J 5.475 1.451 1.628 0.127

Hansen (p- val) 0.019 0.228 0.202 0.722

Anderson-Rubin 4.575 1.632 2.881 0.485A-R (p -val) 0.001 0.167 0.023 0.746

Note:

1. Robust standard errors in brackets. Clustered (by district) in OLS models. *** p<0.01, ** p<0.05, * p<0.1

2. See Notes to Table 3.

OLS - IV

Table 4: Robustness: Killings disaggregated by:

(1) # Civilians killed, (2) # Maoists killed, and (3) # Security Personnel killed

NB - IV

OLS NB ECM OLS NB OLS-IV NB-IV

ln(Total Deaths)t 1 0.0897 0.103 0.909*** ln(Total Deaths)t 1 0.0465 0.104[0.0627] [0.105] [0.062] [0.109] [0.102]

Consumptiont 0.613 1.143 0.635 1.172 Consumptiont 7.831 12.45*[0.572] [0.783] [0.568] [0.783] [7.013] [6.941]

Consumptiont 1 0.917** 2.160*** 1.515* 0.947** 2.213*** Consumptiont 1 2.244 1.099[0.442] [0.719] [0.786] [0.448] [0.716] [7.888] [8.041]

Consumptiont 2 0.00175 0.25 0.0335 0.242 Consumptiont 2 8.859 15.52**[0.467] [0.865] [0.475] [0.887] [6.419] [6.839]

Consumptiont 0.613[0.578]

Consumptiont -1 0.073[0.368]

N 340 340 340 340 340 N 340 340k 77 77 77 76 76 k 77 77

1.143*** 1.152*** 1.11***z CONSUMP 1.556 2.003** - 1.584 2.047** z CONSUMP 2.023** 3.771***

FIRST STAGE:#Instruments 4Kleibergen-Paap (WI) 0.496Hansen's J 0.054Hansen's J (p- value) 0.816Anderson-Rubin (A-R) 3.547Anderson-Rubin (p -val) 0.008

Note: 1. Robust standard errors in brackets. Clustered (by disctrict) in OLS models. *** p<0.01, ** p<0.05, * p<0.1 2. See Notes to Table 1 and Table 3

Table 5: Total Deaths from Maoist incidents and its association with Consumption SpendingDependent Variable : ln(Total Deaths)

Uninstrumented Instrumented

ln(Total Deaths)t 0.0377 0.041 0.05 0.031 0.015 0.020 0.085 0.112 0.113 0.101 0.094 0.091[0.0765] [0.078] [0.078] [0.078] [0.080] [0.078] [0.099] [0.103] [0.101] [0.103] [0.100] [0.097]

Vegetationt 8.955 10.09 12.74 10.33 11.11 7.445 9.682 13.29 17.54 13.88 15.21 8.313[14.51] [14.45] [14.28] [14.28] [14.56] [14.72] [20.31] [20.25] [20.45] [20.23] [20.08] [21.02]

Vegetationt 1 6.765 7.461 7.973 8.333 10.09 3.724 18.99 22.04 21.23 24.58 26.06 4.906[13.13] [13.64] [13.35] [13.46] [13.10] [13.28] [22.24] [22.95] [23.02] [22.95] [22.27] [23.55]

Vegetationt 2 35.49** 39.57*** 37.45** 40.60*** 41.86*** 40.65*** 67.13*** 70.01*** 69.17*** 71.51*** 73.11*** 74.03***[13.93] [14.66] [14.51] [14.49] [15.81] [15.27] [22.53] [23.69] [23.98] [23.59] [24.43] [23.65]

Neighborhood2 0.264** 0.324** 0.326* 0.514***[0.125] [0.131] [0.167] [0.178]

Proportion SC 0.916 1.022 1.443 2.338*[0.704] [0.710] [1.301] [1.345]

Proportion ST 0.841 1.606** 2.101 3.333***[0.832] [0.796] [1.320] [1.286]

Consumption GINI 1.076 1.409 0.952 1.412[1.162] [1.181] [1.467] [1.476]

Value of Mining Output 0.059 0.069 0.082 0.105[0.048] [0.048] [0.087] [0.097]

z ( VEGt ) 2.74*** 3.17*** 3.25*** 3.29*** 3.35*** 2.69*** 3.10*** 3.50*** 3.58*** 3.67*** 3.73*** 2.73***N 340 340 340 335 340 335 340 340 340 335 340 335k 81 81 81 81 81 85 81 81 81 81 81 85

1.067 1.101 1.106 1.122 1.103 1.027#Instruments 4 4 4 4 4 4 4 4 4 4 4 4Kleibergen-Paap 10.26 9.76 10 10.35 9.684 9.409Hansen's J 0.019 0.037 0.056 0.075 0.016 0.003Hansen (p- val) 0.889 0.847 0.812 0.784 0.898 0.957Anderson-Rubin 2.743 3.519 3.453 3.789 3.576 2.845A-R (p -val) 0.029 0.008 0.009 0.005 0.007 0.025Note: 1. Robust standard errors in brackets. Clustered (by district) in OLS models. *** p<0.01, ** p<0.05, * p<0.1 2. See Notes to Table 3.

Table 6: Controls: Spatial Effects, SC/ST, Mining, InequalityDependent Variable : Total Deaths (logged for OLS-IV)

OLS-IV NB-IV

OLS-IV NB-IVln(Total Deaths)t 1 0.054 0.069

[0.061] [0.095]Vegetationt 3.059 5.124

[13.57] [20.75]Vegetationt 1 19.87 42.60*

[13.28] [23.96]Vegetationt 2 39.13** 74.16***

[14.98] [25.84]BIHAR × Vegetationt 7.416 24.24

[8.841] [15.65]BIHAR × Vegetationt 1 9.27 16.77

[7.699] [15.99]BIHAR × Vegetationt 2 7.709 4.382

[12.42] [21.60]CHATTIS × Vegetationt 10.09 21.25

[25.02] [31.57]CHATTIS × Vegetationt 1 4.477 28.92

[18.77] [24.02]CHATTIS × Vegetationt 2 44.87*** 95.39***

[8.697] [19.31]JHAR × Vegetationt 36.51** 83.11***

[17.77] [28.72]JHAR × Vegetationt 1 2.222 2.885

[12.36] [23.21]JHAR × Vegetationt 2 6.546 6.761

[17.08] [28.45]z VEG (Andhra Pradesh) 3.641*** 3.497***z VEG (BIHAR) 2.564*** 3.271***z VEG (CHATTISGARH) 0.244 0.568z VEG (JHARKHAND) 2.775*** 3.969***

N 340 3400.98***

Note:1. z VEG (Andhra Pradesh) is the z- statistic for the hypothesis of the total effect for AP: Vegetationt + Vegetationt 1 + Vegetationt 2 = 0. z VEG (JHARKHAND) is the z- statistic for the total effect for Jharkhand: (Vegetationt + Vegetationt 1 + Vegetationt 2) + (JHAR×Vegetationt + JHAR×Vegetationt 1 + JHAR×Vegetationt 2 ) = 0

Table 7: Robustness: By StateDependent Variable : Total Deaths (logged for OLS-IV)

Table A1. Media Sources for Data Base on Maoist Incidents

State Media Source 1 Andhra Pradesh (E) Indian Express; The Hindu

(V) Eenadu (WS) PTI; IANS

2 Bihar

(E) Indian Express; The Hindu; Times of India (Patna ed.) ; Telegraph (V) Hindustan; Prabhat Khabar (WS) PTI; IANS

3 Chhattisgarh

(E) Indian Express; The Hindu (V) Deshbandhu; Harit Pradesh; Navbharat; Hindustan (WS) PTI; IANS

6 Jharkhand

(E) Indian Express; The Hindu; Times of India(Patna ed.); Telegraph (V) Hindustan; Prabhat Khabar (WS) PTI; IANS

(E): English Language Daily (V): Vernacular/Local Language Daily (WS): Wire Services. PTI: Press Trust of India; IANS: India Abroad News Service (2000-2009).

Mean sd N

Total Deaths (#) overall 11.51 45.44 340

within 28.20

Civilian Deaths (#) overall 4.506 23.78 340

within 17.72

Maoist Deaths (#) overall 4.353 13.27 340

within 8.790

Security Deaths (#) overall 2.649 14.71 340

within 9.828

ln(Total Deaths) overall 1.205 1.401 340

within 0.844

ln(Civilian Deaths) overall 0.686 1.054 340

within 0.711

ln(Maoist Deaths) overall 0.719 1.122 340

within 0.701

ln(Security Deaths) overall 0.430 0.888 340

within 0.605

Vegetation (NDVI measure, range [1, +1]) overall 0.411 0.310 340

within 0.011

Vegetation (Predicted) overall 0.411 0.310 340

within 0.005

Consumption ('000 Rupees per month) overall 0.640 0.243 340

within 0.136

Consumption (Predicted) overall 0.640 0.223 340

within 0.095

Neighborhood2 (number of closest two overall 0.903 0.802 340

districts with killings) within 0.478

Proportion Scheduled Caste (SC) overall 0.187 0.118 340

within 0.075

Proportion Scheduled Tribe (ST) overall 0.117 0.187 340

within 0.059

Consumption Gini overall 0.257 0.068 335

within 0.044

Consumption 90 percentile / 10 percentile overall 0.427 0.110 335

within 0.076

Value of Iron Ore Output (Rupees Billion) overall 0.443 2.942 340

within 1.532

Value of Bauxite Output (Rupees Billion) overall 0.013 0.060 340

within 0.032

Value of Mining Output (sum of iron ore and overall 0.455 2.950 340

bauxite) within 1.542

Note: 2004-2008 for 68 districts in AP, BI, CH, JH.

Table A2: Descriptive Statistics

This portion not for review WEB APPENDIX 1: CASI DATABASE ON MAOIST INCIDENTS Existing Data Sources The official source on Maoist-related violence is released by the Ministry of Home Affairs of the Government of India. This data are divided into two categories: number of casualties (deaths, disaggregated by civilian, security forces and Maoist) and the number of violent incidents. However, these data are only available at the state level. Other popularly used data sources are the RAND-MITP Terrorism Incident database; data from the Worldwide Incidents Tracking System (WITS) from National Counter Terrorism Centre and data from the South Asian Terrorism Portal (SATP). The Rand-MIPT and the WITS data sets are world-wide and their Indian data is ad hoc and do a poor job at capturing Indian data reported by the non-English language press. The SATP data set has been assembled by the Institute for Conflict Management, available at their website, the South Asia Terrorism Portal. These data are based primarily on reports in the major English press, but do a considerably better job than the other two, especially in more recent years. However, this data set is only from 2005 and even after that it is not comprehensive. The CASI database Our Center for the Advanced Study of India (CASI) data set goes back to 2000. It has been compiled from multiple media sources. Even though two national English dailies – the Indian Express and The Hindu – have covered the Naxal issue extensively, they provide only partial coverage. We drew on ten other media sources: two additional English-languge newspapers: Times of India (Patna edition) and the Telegraph; six regional language press sources: Eenadu; Hindustan; Prabhat Khabar, Deshbandhu; Harit Pradesh; Navbharat; and two wire services: PTI and IANS. We had to overcome a number of barriers to the creation of a reliable data set – access to available sources, cost of accessing data and the quality of data sources. Access Few newspapers have digitized archives that are freely available on the internet. Where available, we thoroughly documented that data. The remaining data were had to be acquired by contacting the various newspapers for access to their archives. We used a combination of methods. Our researchers contacted media offices directly. Since Maoist violence is a politically extremely sensitive topic locally, some media offices were understandably apprehensive about our request and others cited prohibitively high user fees. To overcome this hurdle, we fell back on the networks between the Indian academic community and senior, influential media persons to be able to secure access. Regional newspapers were more forthcoming and prompt in granting access as opposed to national dailies (both English and Hindi).

Another problem faced by researchers was where newspapers had handed over the management of their archives to private companies specializing in maintaining the documents of large companies. In that case, the agent company decided the terms and conditions of access, which usually came at prohibitive cost. We were able to obtain access only after intervention by editors and some protracted discussions between researchers, the media house and the company where the archives were stored. Cost Some media houses charge a daily user fee, apart from fees for photocopying the relevant material. The cost of accessing archives managed by private documentation companies is very high. Fees are charged under various heads – a daily user fee, a fee per issue of newspaper viewed, a fee per monthly file viewed, and 20 per cent additional charge in lieu of the storage charges being paid by the media house to the private company. Suffice it to say that the data and field work come at a significant cost. We believe the data to be far superior to the sources mentioned in the first paragraph. Quality In most cases newspapers were stored in poor conditions, with issues for an entire month often bound into a single file for the purpose of storage. In most of the archives, the files were not stored in any systematic manner – for instance year-wise – that could ensure easy access. A lot of effort was spent re-ordering the files. Since files were sometimes missing relying on a single source for local data would compromise the quality of the data. We have therefore followed the more costly strategy of scouring multiple sources for each state. We therefore have attempted to ensure as much overlapping coverage as possible, making the data far more reliable than any other data source. Uneven Reporting Apart from access, reportage of incidents is uneven. The fact that different incidents were reported on the same day in different newspapers implies that several incidents are not being reported, and hence, considerable underreporting of the problem by any individual newspaper. Coverage in national newspapers is stronger in some regions than others. For some years the overlapping coverage of incidents in two national English-language newspapers – The Hindu and Indian Express – was less than a fifth, while the two combined were a small fraction of the total number of incidents in the final CASI database, since it was built from multiple sources. Coding For each incident we have intended to record the following: - date (month and year) - location (district and village) - type of incidence (bomb explosion, kidnapping etc…) - number of Maoists killed in that incident - number of non-Maoist civilians killed - number of security personnel killed - other information relevant to the incident

While different members of our team accessed news sources from different parts of the country, a copy of each story was sent to a core team for coding to ensure consistency. The exception was Eenadu which needed a person fluent in Telugu as well as with a strong social science background so that the coding could be done directly. Even this painstaking effort is not without problems. When several newspapers report the same incident, the news is sometimes printed on different dates. News reports often carry different placelines (place of reportage), as well as different accounts of the location of the incident. Details on the number of casualties and injuries, how many of those involved were Maoist, how many were civilians and how many were security forces have different figures in different accounts. In some stories, the number of injured and dead remains unspecified. Unless an incident of Maoist violence is large in scale or intensity, it is rarely followed up. A frequently used phrase during reportage is ‘suspected Maoists’ and ‘Maoist sympathizers’. In doubtful cases, where during the time of reportage, the journalist is unsure about details and the discrepancies are rarely resolved through follow-up articles. We have coded them as civilians. Another issue relates to the Salwa Judum, a militia consisting of local tribal youth organized by the government of Chhattisgarh to counter the Naxalite violence in the region. Since members of this militia had the status of Special Police Officers (SPOs) and received training and wages from the Chhattisgarh state government we have coded them as security forces (A judgment of the Indian Supreme Court in 2011 declared this militia as illegal). Our intention in providing these details is to indicate that data pertaining to Maoist casualties is not problem-free, but we have taken as much care as possible to present the most representative picture to date. All politics is local, and so is all political violence. The Rand-MIPT and the WITS data are selected data from large newspapers, and considerably underreport the actual extent of the conflict. When an event is newsworthy, it is better covered than if it is confined locally. For an issue with enormous policy economic and political implications as the Maoist conflict, we think the data must be up to the task. We think ours is.